Data Science Seminar

Spectral clustering techniques are valuable tools in signal processing and machine learning for extracting a far more varied collection of cluster structures than the popular k-means clustering algorithm. The effectiveness of spectral clustering stems from constructing a non-linear embedding based on creating a similarity graph and computing the spectral decomposition of the Laplacian matrix. However, spectral clustering methods fail to scale to large data sets because of high computational cost and memory footprint. Hence, the scalability of spectral methods is of increasing importance in modern applications, and several research efforts have been conducted in this regard. This talk will review related work on accelerating spectral clustering, discussing merits and limitations. The second focus in this talk is to present a rigorous algorithm for regulating accuracy-efficiency trade-offs that builds on the Nystrom approximation, a sampling-based method for computing low-rank approximations to large positive semi-definite matrices. We present two new theoretical results to understand the influence of leveraging low-rank approximations in the context of spectral clustering, supported by numerical experiments.

Bio: Farhad Pourkamali-Anaraki is an assistant professor in the Department of Computer Science at the University of Massachusetts Lowell. He received his PhD in Electrical Engineering from the University of Colorado Boulder in 2017. He received the Best Paper Award at the 19th IEEE International Conference on Machine Learning and Applications (ICMLA). His research lies at the intersection of statistics, optimization, and machine learning on the one hand and the intersection of theory, applications, and implementations on the other side. In particular, he works on extending the use of machine learning algorithms to other domains, including reliability analysis of critical infrastructures in Civil Engineering and advanced manufacturing in Mechanical Engineering.

The use of deep neural networks has become increasingly popular for computer vision tasks. These models have the potential to achieve great accuracy in recognizing images, but they are often called "black boxes" because they generally suffer from a lack of interpretability. In fact, evidence has emerged, where some deep learning methods appeared to perform well, but based their decisions on confounding rather than truly relevant information. In this talk, I will present my work in developing a prototypical case-based interpretable neural network known as a prototypical part network (ProtoPNet), which classifies an input image by comparing various parts of the image with learned prototypical features known as prototypes. I will also present the use of a prototypical case-based interpretable neural network model for analyzing medical images, with a focus on mammograms. Such an interpretable model will automatically detect areas of anomaly (such as spiculated mass margins) and present prototypical examples of similar anomalies found in other patients as explanations for their predictions. In particular, I will discuss a human-in-the-loop training scheme, which is based on a small set of doctor-annotated images and is specifically designed to prevent the model from using confounding information. The explanations generated by our interpretable neural network for analyzing mammograms thus augments the reports generated by radiologists, and may be able to assist doctors in diagnosing patients in the future.

Bio: Dr. Chen is an Assistant Professor of Computer Science at the University of Maine. His research involves the design of interpretable machine learning models that can be understood (“interpreted”) by human beings. In particular, Dr. Chen is interested in developing new techniques to enhance the interpretability and transparency of machine learning models, especially deep learning models, and in applying such techniques to healthcare, finance, and other application domains, where high-stakes decisions are made and interpretability is key for whether one can trust the predictions made by machine learning models. Dr. Chen also participates in the Maine environmental DNA (eDNA) project: he is developing computational techniques to understand eDNA.

Game-tree search is an important field in AI, with numerous applications. The centrality of game trees is made evident by the thousands of different search algorithms found in the literature for handling them. Nearly all such algorithms have hard-wired decision-making w.r.t. nodes they expand, heuristics the compute, or even learning they perform en-passant. Notable exceptions are algorithms that reason explicitly about the tradeoff between computational resource consumption and solution quality, most especially the seminal work on rational metareasoning by Russell and Wefald from circa 1990. Rational metareasoning (AKA "the right thing") is extremely hard to apply to actual search, thus has taken a long time to bear fruit.

This talk examines some of the reasons for this difficulty, and presents schemes to bypass the difficulties by using properties of submodular functions, and by relaxing myopic assumptions made by Russell and Wefald. Although the results are rather general, we also briefly discuss their application to improving Monte-Carlo tree search.

Domain-independent, classical planners require symbolic environment models as input, exhibiting the so-called "Knowledge Acquisition Bottleneck," the common bottleneck of symbolic AI systems that the human must encode the model by hand. Meanwhile, although Deep Learning (DL) has achieved significant success in many fields, the knowledge is encoded in a subsymbolic representation which is incompatible with symbolic systems such as planners.

In AAAI18, we proposed LatPlan, an unsupervised architecture combining deep neural networks and classical planners. Given a pair of images representing the initial and the goal states (planning inputs) and an unlabeled dataset of image pairs showing a subset of transitions allowed in the environment (training inputs), LatPlan automatically converts the input into a discrete, classical planning problem, solves it with a classical planner, then returns a visualized plan execution. We evaluated LatPlan using image-based versions of 3 planning domains: 8-puzzle, Towers of Hanoi and LightsOut.

I will discuss the extension of the system that we published in two more conference papers, an arxiv paper and another paper under submission. If time permits, I will also discuss the prospects of neural-symbolic systems, including other work performed by IBM Research Cambridge (MIT-IBM Watson AI Lab).

Bio: Masataro Asai received his Ph.D from the University of Tokyo in 2018 and has been at the IBM Research Cambridge (MIT-IBM Watson AI Lab) since last summer. His original field of study is Classical Planning. His hobby is competitive driving, car modification, and computer hacking in Lisp. He is an author of Numcl, a Numpy clone written in Common Lisp.